面向城市消防站选址规划的时空动态火灾风险建模分析

发布时间：2017-12-28 20:15

本文关键词：面向城市消防站选址规划的时空动态火灾风险建模分析　出处：《中国科学技术大学》2017年博士论文　论文类型：学位论文

【摘要】：当前,中国的城镇化建设处于高速发展的阶段,与此同时,城市火灾风险居高不下。我国城市的消防规划相对滞后于城镇化的快速发展,特别对于城市的未来发展以及对火灾风险的发展考虑得不够全面,同时,考虑时空动态分布的火灾风险影响下的消防资源布局方面的研究也较少。本文通过引入相关数学模型来综合考虑时空变化对城市火灾发生的影响,充分考虑火灾发生的时空变化特征,并在此基础上发展一种与城市发展相适应的、动态的消防站选址规划模型,可为提高城市消防规划的科学性提供理论支撑。主要工作如下:1.城市尺度火灾发生的空间分布研究通过使用机器学习算法和空间计量经济学模型对市域尺度的城市火灾的发生进行空间建模,从而预测或者解释火灾发生的空间分布。首先,我们使用机器学习算法进行分析。针对特征选择得到的重采样结果显示,在训练集中,四种机器学习算法中随机森林模型(RF)的预测精度最高。样本点距离消防站的距离、人口的分布、企业工厂的分布、道路密度、温度的分布以及海拔是选入RF中的自变量,说明这些变量对火灾风险的贡献度最大。同时随机森林算法在训练集和测试集上均预测效果较好,差别不大,说明其较为稳健。之后,通过对空间计量经济学模型进行分析,结果显示空间通用模型(SAC)的赤池信息准则(AIC)值最小,也即拟合效果最好。结果显示道路密度对火灾密度有正相关的关系,而温度、海拔以及距离消防站的距离这三个变量均与火灾密度呈负相关的关系。结果说明人类的活动对于火灾发生有着相当重要的作用,同时也对空间回归模型的贡献度相对较大。同时,SAC模型在训练集上均表现良好,但在测试集上均与训练集的预测效果相差较大。这说明空间通用模型更适合对火灾风险进行解释。总之,在模型的适用性方面,我们发现RF能够表现出更稳健的预测能力,鲁棒性较强。而对于空间计量经济学模型,它们则能够从特定数据集(训练集)中隐藏的空间分布规律和聚类特征入手,采用空间权重矩阵准确地刻画市域尺度下城市火灾发生的空间规律与影响较大的自变量,更适合于解释火灾的发生,但总体上要强于随机森林算法。同时,通过对最终得到的SAC与RF模型在整个数据集上进行预测,结果显示,SAC在总体上要优于RF。进一步通过对模型的残差进行分析和可视化,结果显示SAC比RF更能解释火灾发生的空间结构。2.城市尺度火灾发生的时间分布研究通过使用小波变换的方法对火灾与气象因素的时间序列进行分析,进而揭示时间序列中存在的潜在规律与变化共性,之后使用经典的自回归移动平均模型与MCMC对火灾频次的时间序列进行拟合,最终得到时间维上的城市火灾风险发展趋势。首先,通过小波变换来分析火灾频次的时间序列,结果显示存在四个特征时间尺度,分别为4,18,34,56个月,火灾频次的周期性波动在56个月(4-5年)这样的时间尺度上周期性最强,适合于在较大的年际尺度上进行火灾周期性的研究分析。可以从定性的结果中推测在2010-2011年火灾频次将呈现下降趋势,之后将继续上升。其次,气象因素的时间序列通过小波分析,发现不同的变量存在不同的特征时间尺度集合。它们均包含18个月的特征时间尺度,这说明气象因素也和火灾频次的时间序列拥有共同的平均变化周期为12个月。最后,通过使用经典的时间序列模型对火灾频次进行建模,通过比较,最终选用ARIMA(0,1,1)模型来进行建模。另一方面,我们选用MCMC模型来进行时间序列建模,结果显示日照时长的系数值最大,表示其对火灾发生的影响最大。同时,日照时长与相对湿度的系数值为正,表示与火灾发生的频次正相关;平均温度与降水量的系数值为负,表示与火灾发生的频次负相关。另外,去除气象变量后的MCMC模型与包含气象变量的MCMC模型进行了对比,结果表明加入气象因素的MCMC模型的拟合度更好,更能勾勒时间序列中的波动。3.城市尺度火灾发生的时空分布研究通过使用以最小二乘法(OLS)为基础的全局线性回归模型(LM)进行分析,对比了全局模型与局部模型在火灾风险预测方面的效果,最终可以得到火灾风险分布的时空模型。首先,我们使用空间交叉验证(SCV)对自变量进行选择。使用SCV进行线性回归训练的结果显示,道路密度与工厂企业的空间分布是火灾风险建模中最重要的变量,与火灾的发生密切相关,但是前者对于火灾风险正相关,而后者为负相关。之后,我们使用选择的自变量进行了地理加权回归(GWR)与时空加权回归(GTWR)两种局部拟合模型对火灾风险进行解释与预测。结果显示,拟合效果GTWRGWRLM。通过提取模型中的自变量的系数并进行可视化,结果显示,对于相同的自变量,其系数值与显著度水平在不同时间截面的相同位置处发生了一定程度的变化。最后,通过对不同区域显著变量的数目进行分析,结果显示参数的空间异质性分布规律主要集中在城市的非中心区,往往是一些人类局部活动性较为集中的次中心地带或者集镇、县城等地区。通过对三个模型产生的残差进行半异方差检测,结果显示GTWR对于火灾发生的空间结构的解释程度最强。同时,在模型的独立测试阶段,GTWR也显示出比GWR、OLS更为准确的预测能力。4.基于火灾发生时空动态模型的城市消防站选址规划模型首先,在前文工作的基础上,我们可以预测未来一段时间的火灾风险空间分布,进而建立改进的城市消防站选址规划模型。通过使用遗传算法进行选址模型的模拟仿真,结果显示使用基于匹配度与多重部分覆盖模型的选址模型更能体现出实际消防规划中的"等级对应"思想,也即不同等级的风险应和同等级的预防措施相对应。风险越高,消防站的出动时间应越小,救援路径应越短。因此,当目标函数为匹配度时,我们所建立的新模型与其他经典选址模型(P-中值,最大覆盖)相比,目标值最大且更加合理。另外,我们使用基于容量限制的路径寻优算法来模拟当发生并发火灾时,消防车对于火点进行扑救的路径选择问题。结果显示,基于容量限制的最短路径寻优算法CCRP可以解决消防车在基于城市道路网进行扑救时的最短路径选择问题。同时,这也可以模拟将空间节点上的人群进行群体疏散时来选择最短路径的情形。
[Abstract]:At present, the construction of urbanization in China is at a stage of rapid development. At the same time, the risk of urban fire is high. Our country city fire planning lagging behind the rapid development in urbanization, especially for the future development of the city and the development of fire risk are not considered comprehensively, at the same time, considering the fire risk of fire resource layout effect of temporal and spatial dynamic distribution under the less. In this paper, through introducing the relevant mathematical model considering the influence of temporal and spatial variation of fire in the city, fully considering the characteristics of temporal and spatial variation of fire, and on the basis of the development of a city, and adapt to the development of dynamic fire station location planning model, which can provide theoretical support for improving the scientific nature of city fire protection planning. The main works are as follows: 1., the spatial distribution of fires in urban scale is studied by using machine learning algorithm and spatial econometrics model to model the occurrence of urban fires in city scale, so as to predict or explain the spatial distribution of fires. First, we use machine learning algorithms to analyze. The resampling results obtained from the feature selection show that the prediction accuracy of the random forest model (RF) is the highest among the four machine learning algorithms in the training set. The distance from the sample point to the fire station, the distribution of the population, the distribution of the plant, the density of roads, the distribution of temperature and altitude are the independent variables selected into RF, which indicate that these variables contribute most to the fire risk. At the same time, the prediction effect of random forest algorithm on the training set and test set is good, and the difference is not significant, which shows that the algorithm is more robust. Then, through the analysis of spatial econometrics model, the results show that the AIC of SAC is the smallest, that is, the best fitting effect. The results showed that road density had a positive correlation with fire density, while the three variables of temperature, altitude and distance from fire station were negatively correlated with fire density. The results show that human activities play an important role in the occurrence of fire, and also have a relatively large contribution to the spatial regression model. At the same time, the SAC model shows good performance on the training set, but the prediction effect of the test set is quite different from the training set. This shows that the spatial general-purpose model is more suitable for the interpretation of fire risk. In general, in terms of the applicability of the model, we find that RF can show more robust predictive ability and stronger robustness. The spatial econometric model, they can from a specific data set (training set) hidden in spatial distribution and clustering of the spatial weight matrix to accurately describe the variables of large space law and influence of city fire occurred under the city scale, more suitable for explaining the occurrence of fire, but on the whole. In the random forest algorithm. At the same time, the final SAC and RF models are predicted on the whole data set, and the results show that SAC is better than RF in general. Through the analysis and visualization of the residual error of the model, the results show that SAC can explain the spatial structure of fire more than RF. Through the analysis method of wavelet transform in time series of fire and meteorological factors of 2. city scale fire time distribution research, which reveals the potential rule and change in common in time series, after the use of classical autoregressive moving average time series of fire frequency model and MCMC fitting, finally get the development trend of city the time dimension of the fire risk. First of all, through the wavelet transform to time series analysis of fire frequency, the results show that there are four characteristic time scales, respectively, for 4,18,34,56 months, the fire frequency of periodic fluctuation in 56 months (4-5 years) so that on the time scale of the most cyclical, suitable for the analysis of periodic fire in large interannual scale the. It can be deduced from the qualitative results that the frequency of fire will decline in 2010-2011 years and will continue to rise. Secondly, the time series of meteorological factors is analyzed by wavelet analysis, and it is found that different variables have different characteristic time scale sets. They all contain 18 months of the characteristic time scale, which shows that the meteorological factors and the time series of the frequency of the fire have a common average change period of 12 months. Finally, by using the classical time series model to model the frequency of the fire, the ARIMA (0,1,1) model is used to model the model by comparison. On the other hand, we use the MCMC model to model the time series. The results show that the maximum length of the sunshine is the largest, which indicates that it has the greatest impact on the fire. Meanwhile, the relationship between sunshine duration and relative humidity is positive, indicating a positive correlation with the frequency of fire occurrence. The average temperature and precipitation are negative, indicating a negative correlation with the frequency of fire occurrence. In addition, the MCMC model which removes the meteorological variables is compared with the MCMC model including the meteorological variables. The results show that the MCMC model with meteorological factors is better fitted and better able to describe the fluctuation in the time series. Study on the temporal and spatial distribution of 3. city scale fires by using the least squares method (OLS) based global linear regression model (LM) analysis, comparison of the global model and local model in fire risk prediction effect, finally can get the spatio-temporal model of fire risk distribution. First, we use the space cross validation (SCV) to select the independent variables. The results of linear regression training using SCV show that the spatial distribution of road density and factory enterprises is the most important variable in fire risk modeling, and is closely related to the occurrence of fire. However, the former is positively related to fire risk while the latter is negatively correlated. After that, we use the selected independent variables to carry out the GWR.
【学位授予单位】：中国科学技术大学
【学位级别】：博士
【学位授予年份】：2017
【分类号】：TU998.1

【相似文献】